Improved support vector machines model based on multi-spectral parameters

Abstract

To solve the classification accuracy instability problem of support vector machines (SVM) model caused by using original hyper-spectral data or a single spectral feature as modeling feature vector, an improved SVM model with multi-spectral parameters and joint kernel function is proposed in this paper. The feature vector of SVM model is built by multi-spectral parameters using feature selection method based on information quantity and between-class separability and feature extraction method based on minimum noise fraction. The kernel function of improved SVM model is optimized with linear combination of polynomial kernel function and radial basis kernel function to increase the learn ability and generalization ability. And the multi-class classification strategies based on an improved directed acyclic graph is presented in the proposed method. The airborne hyperspectral remote sensing images collected by pushbroom hyperspectral imager (PHI) and airborne visible infrared imaging spectrometer (AVIRIS) are applied to analyze and evaluate the performance of the proposed method in this paper. The experiment results show that the classification accuracy is better than 90% and the plant fine-classification ability for small sample and similar spectral features is realized.